Skip to content
Get packs

Cosmos Policy Evaluation

Set up and run NVIDIA Cosmos Policy robot evaluations on LIBERO and RoboCasa simulators

Source Download skill
Repository
Orchestra-Research/AI-Research-SKILLs
Stars
14,195
Licence
MIT
Files
3 files

What it does

This skill gives an AI agent step-by-step instructions for benchmarking NVIDIA's Cosmos Policy, a robot-control model, inside two simulated environments (LIBERO and RoboCasa). It covers setting up a blank GPU machine, configuring headless rendering so simulations run without a screen, running a quick smoke test, parsing the result logs, and scaling up to full benchmark runs. Without it, the agent would have to work out these commands and flags from scratch.

Great for

  • Robotics researchers who want to reproduce Cosmos Policy results on LIBERO or RoboCasa
  • ML engineers setting up headless GPU evaluation on a cluster
  • Teams profiling inference latency of a robot policy on local or cluster GPUs

Skip it if

  • Training or fine-tuning a model from scratch: the skill itself says to use the official Cosmos training docs instead
  • Real-robot testing: the skill covers simulation only, and running commands against physical hardware is outside its scope
  • Anyone without a suitable NVIDIA GPU (roughly 16-18 GB VRAM) and the time for multi-hour runs, since the full benchmarks take hours

Try asking

  • “Set up cosmos-policy on a blank GPU machine and run a LIBERO smoke evaluation”
  • “Configure EGL headless rendering and run the RoboCasa single-task evaluation”
  • “Run the full LIBERO benchmark across all four task suites with 50 trials”
  • “Parse the latest LIBERO evaluation JSON log and summarise the results”
Step-by-step guide: how to use Cosmos Policy EvaluationInstall, first run, examples, tips and FAQs

Install in 3 steps

  1. 1

    Download the skill

    The ZIP contains the 3 original files, the MIT licence and an attribution note.

    cosmos-policy.zip
  2. 2

    Add it to your agent

    Claude Code (all projects)

    mkdir -p ~/.claude/skills && unzip cosmos-policy.zip -d ~/.claude/skills/

    Claude Code / Agent SDK (this project only)

    mkdir -p .claude/skills && unzip cosmos-policy.zip -d .claude/skills/

    Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.

  3. 3

    Ask for the task

    No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.

Get better results

  • Run the 1-trial LIBERO smoke test first to confirm the environment works before launching the multi-hour full benchmark.
  • Review the commands before running them: they execute code, install packages from git URLs, and download model checkpoints.
  • Check the log directory output paths and GPU settings (CUDA_VISIBLE_DEVICES, MUJOCO_EGL_DEVICE_ID) match your machine before starting long jobs.

What's inside

  • references/libero-commands.md
  • references/robocasa-commands.md
  • SKILL.md

Redistributed unmodified from Orchestra-Research/AI-Research-SKILLs under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.